Skip to content

Mom, these are not the bayesian causal forests I wanted! Even that weirdo girl in math class got cooler ones. Why are you always like this?

Latest

Choose a tag to compare

@Gattocrucco Gattocrucco released this 03 Nov 22:28
· 17 commits to main since this release
Immutable release. Only release title and notes can be modified.
v0.14.0
f54d017
  • Polish the bcf submodule; main things to know:
    • General improvements in speed, memory, error checking, and consistency
    • New maxdepth_mu and maxdepth_tau parameters
    • In predict_potential_outcomes, now rho defaults to 0.0 instead of 0.5, and it can be negative
    • With adaptive coding, the prior on the coding parameters b0 and b1 excludes the slice $|b|<0.01$ for technical numerical reasons. This does not change the model in practice because those parameters are not identified.
    • Save to file the full model, not just the parts required to compute predictions
  • Improve saving/loading models to file
    • Guarantee backwards compatibility: new bartz versions will always load files created by older versions
    • New methods Bart.save_npz and Bart.load_npz, mirroring the bcf ones
      • Deprecate Bart.dump and Bart.load
    • Much smaller file sizes in general
  • Improvements in run_mcmc callbacks and logging
    • Main feature: the progress bar now shows "tree life", which measures how long lived the trees are in the MCMC. Short-lived trees mean the MCMC is moving fast, long-lived trees mean the MCMC is partially stuck.
  • Fix numerical accuracy and stability problems in the variable selection feature
    • This should not make a difference in practice, but it makes the MCMC more robust to corner cases and makes it possible to do precise prior comparison checks
  • Improvements with multivariate outcomes
    • About 2x faster on CPU on small datasets with 2-4 outcomes
    • More accurate posterior when the outcomes are very correlated, should not make a difference in most cases
  • Fixes in Bart and mc_gbart
    • Raise on NaN in x_train and x_test to avoid the pitfall of users assuming NaN means "missing value". Missing values in predictors are not supported.
    • Fix that Bart.predict would not work with pandas dataframes
    • In mc_gbart, the sigest OLS degrees of freedom now count the intercept ($n - p - 1$ instead of $n - p$)
  • Generally faster, mostly on CPU
  • Rewrite prior sampler; the new interface makes it easier to match the setup of a given MCMC